riturajtiwari-ai/emotion_model

Model Card for Emotion Detection CNN

Model Details

Model Description

This is a lightweight Convolutional Neural Network (CNN) for facial expression / emotion classification from grayscale face images. The uploaded model is saved in Keras HDF5 (.h5) format and expects images resized to 48 × 48 × 1. It outputs probabilities across 7 emotion classes using a final Softmax layer.

The H5 file contains the model architecture, trained weights, optimizer weights, and training configuration. The exact class-label order is not stored inside the H5 file, so the label order should be verified from the original training script before publishing or deploying.

• Developed by: Rituraj Tiwari
• Funded by [optional]: Not applicable / self-project
• Shared by [optional]: Rituraj Tiwari
• Model type: Keras Sequential CNN image classifier
• Task: Facial expression / emotion recognition
• Input: Grayscale face image, 48 × 48 × 1
• Output: 7-class emotion probability vector
• Language(s) (NLP): Not applicable; this is a computer vision model
• License: MIT
• Finetuned from model [optional]: Not applicable; the H5 architecture appears to be a custom CNN, not a pretrained backbone

Model Sources [optional]

• Repository: Add your GitHub repository link here
• Paper [optional]: Not applicable
• Demo [optional]: Add your live demo or Hugging Face Space link here

Uses

Direct Use

This model can be used to classify a detected face image into one of seven facial expression categories. It is intended for educational projects, demos, portfolio projects, and basic emotion-recognition experiments.

Expected pipeline:

1. Detect or crop a face from an image/video frame.
2. Convert the face crop to grayscale.
3. Resize it to 48 × 48.
4. Normalize pixel values, commonly by dividing by 255.0.
5. Run inference using the Keras model.
6. Map the highest-probability output index to the correct emotion label.

Common emotion labels for this type of project are:

["Angry", "Disgust", "Fear", "Happy", "Sad", "Surprise", "Neutral"]

Important: the H5 file confirms there are 7 output classes, but it does not store the class names or their order. Confirm the exact label order from the original dataset loader or training script.

Downstream Use [optional]

This model can be integrated into:

• A Flask, FastAPI, Django, or Streamlit web application
• A webcam-based real-time emotion detection system
• A React frontend with a Python backend
• A student project or academic demonstration
• A basic human-computer interaction prototype

Out-of-Scope Use

This model should not be used for:

• Medical, psychological, legal, hiring, policing, or high-stakes decision-making
• Surveillance or monitoring without user consent
• Determining a person's true internal emotional state
• Identity recognition or face verification
• Production systems where incorrect predictions could harm users

Facial expression recognition is uncertain and context-dependent. A facial expression does not always represent a person's real emotion.

Bias, Risks, and Limitations

The model may perform poorly when the input face image differs from the training data. Possible limitations include:

• Low accuracy in poor lighting, blur, occlusion, extreme face angles, or low-resolution images
• Bias across age groups, skin tones, genders, cultures, facial accessories, and camera conditions
• Confusion between visually similar expressions such as fear/surprise or sad/neutral
• Weak generalization outside the original dataset distribution
• Predictions that reflect visible facial expression, not a guaranteed emotional state

The model uses a relatively small custom CNN architecture, so it may be less accurate than larger modern architectures trained on larger and more diverse datasets.

Recommendations

Users should:

• Validate the model on their own test images before deployment
• Show prediction confidence instead of only the top label
• Avoid using the model for sensitive or high-stakes decisions
• Add a disclaimer when showing predictions to end users
• Confirm the exact class-label order from the training script
• Consider retraining or fine-tuning with a more diverse dataset for real-world use

How to Get Started with the Model

Use the code below to load the model and run prediction on a face image.

import cv2
import numpy as np
import tensorflow as tf

# Load model
model = tf.keras.models.load_model("emotion_model.h5", compile=False)

# Confirm this order from your training script before publishing
class_names = ["Angry", "Disgust", "Fear", "Happy", "Sad", "Surprise", "Neutral"]

# Load a face image in grayscale
img = cv2.imread("face.jpg", cv2.IMREAD_GRAYSCALE)

# Preprocess
img = cv2.resize(img, (48, 48))
img = img.astype("float32") / 255.0
img = np.expand_dims(img, axis=-1)  # shape: (48, 48, 1)
img = np.expand_dims(img, axis=0)   # shape: (1, 48, 48, 1)

# Predict
predictions = model.predict(img)[0]
predicted_index = int(np.argmax(predictions))
predicted_label = class_names[predicted_index]
confidence = float(predictions[predicted_index])

print("Prediction:", predicted_label)
print("Confidence:", confidence)

Training Details

Training Data

The training dataset is not stored inside the H5 file. Based on the model input shape and 7-class output design, this model is suitable for FER2013-style facial expression datasets, where grayscale face images are commonly resized to 48 × 48.

If this model was trained on FER2013 or a similar dataset, document the dataset here:

• Dataset name: Add exact dataset name here
• Dataset source: Add dataset link here
• Number of classes: 7
• Image format: Grayscale face images
• Image size: 48 × 48
• Labels: Confirm exact labels and class order from the training script

Training Procedure

The H5 file includes the following training configuration:

• Optimizer: Adam
• Learning rate: 0.001
• Loss function: Categorical cross-entropy
• Metric: Categorical accuracy
• Backend: TensorFlow
• Keras version saved in file: 2.10.0

Preprocessing [optional]

Recommended preprocessing for this model:

1. Convert image to grayscale.
2. Detect/crop the face region.
3. Resize to 48 × 48.
4. Convert to float32.
5. Normalize pixel values to [0, 1].
6. Reshape to (1, 48, 48, 1) for single-image inference.

Training Hyperparameters

• Training regime: Not stored in the H5 file
• Optimizer: Adam
• Learning rate: 0.001
• Loss: Categorical cross-entropy
• Dropout: 0.5
• Batch size: Not stored in the H5 file
• Epochs: Not stored in the H5 file
• Precision: Not stored in the H5 file; likely standard float32

Speeds, Sizes, Times [optional]

• Model file size: Approximately 9.65 MB
• Trainable parameters: 839,047
• Input shape: (48, 48, 1)
• Output shape: (7,)
• Training time: Not stored in the H5 file
• Checkpoint format: Keras HDF5 .h5

Evaluation

Testing Data, Factors & Metrics

Testing Data

The test dataset is not stored inside the H5 file. Add the test split or validation dataset details here.

Factors

Recommended evaluation factors:

• Lighting condition
• Face angle
• Image quality
• Age group
• Skin tone
• Gender presentation
• Occlusion such as glasses, mask, hand, or hair
• Real-time webcam frames vs. clean dataset images

Metrics

Recommended metrics:

• Accuracy
• Precision
• Recall
• F1-score
• Confusion matrix
• Per-class accuracy
• Inference latency

Summary

This model is a compact CNN designed for 7-class facial expression classification. It is suitable for demos and educational use, but it should be evaluated carefully before real-world deployment.

Technical Specifications

Model Architecture and Objective

The model is a Keras Sequential CNN for multi-class image classification. It uses convolutional layers for feature extraction, max-pooling layers for spatial downsampling, a dense layer for classification features, dropout for regularization, and a final Softmax layer for 7-class probability output.

Layer	Configuration	Output Shape	Parameters
InputLayer	48 × 48 × 1 grayscale image	(None, 48, 48, 1)	0
Conv2D	32 filters, 3 × 3, ReLU, valid padding	(None, 46, 46, 32)	320
MaxPooling2D	Pool size 2 × 2	(None, 23, 23, 32)	0
Conv2D	64 filters, 3 × 3, ReLU, valid padding	(None, 21, 21, 64)	18,496
MaxPooling2D	Pool size 2 × 2	(None, 10, 10, 64)	0
Flatten	Flatten feature maps	(None, 6400)	0
Dense	128 units, ReLU	(None, 128)	819,328
Dropout	Rate 0.5	(None, 128)	0
Dense	7 units, Softmax	(None, 7)	903

Total trainable parameters: 839,047

Software

The uploaded H5 file contains:

• Backend: TensorFlow
• Keras version: 2.10.0
• Model format: Keras HDF5 .h5

For best compatibility, use TensorFlow/Keras 2.x or load with:

tf.keras.models.load_model("emotion_model.h5", compile=False)

Citation [optional]

No citation is available for this model. Add a citation here if you publish a paper, blog post, or project report.

BibTeX:

@misc{emotion_detection_cnn,
  title = {Emotion Detection CNN},
  author = {Rituraj Tiwari},
  year = {2026},
  note = {Keras CNN model for facial expression recognition}
}

APA:

Tiwari, R. (2026). Emotion Detection CNN: A Keras model for facial expression recognition.

Glossary [optional]

• CNN: Convolutional Neural Network, a neural network commonly used for image tasks.
• Softmax: A function that converts raw model outputs into class probabilities.
• Categorical cross-entropy: A common loss function for multi-class classification.
• Dropout: A regularization method that randomly disables neurons during training to reduce overfitting.
• H5 / HDF5: A file format often used to save Keras models.

More Information [optional]

Add links to your project repository, demo, training notebook, dataset, or report here.

Model Card Authors [optional]

Rituraj Tiwari

Model Card Contact

Add your preferred contact email, GitHub profile, or Hugging Face profile here.